dolfin team mailing list archive

[Johan Hake <hake@xxxxxxxxx>]

On Tuesday 30 December 2008 11:39:19 Joachim B Haga wrote:
> Johan Hake <johan.hake@xxxxxxxxx> writes:
> > I have done that several times with no luck. I have tried valgrind, with
> > no directly usefull output (found some MPI memory leaks though) and also
> > tried
>
> I was curious about this, so I ran it myself. And it does apparently detect
> this error:
>
> ==26837== Invalid write of size 8
> ==26837==    at 0x734B727: dolfin::Function::interpolate(double*,
> [...]
> ==26837==  Address 0x4a8b910 is 0 bytes after a block of size 24 alloc'd
> ==26837==    at 0x402209E: operator new[](unsigned)
> (vg_replace_malloc.c:268) ==26837==    by 0x73449B8:
> dolfin::FunctionSpace::Scratch::init(dolfin::FiniteElement (no line numbers
> inside dolfin, maybe add -g to the compile flags?)

Aha! There it was! 

> But it may be hidden in the noise unless you suppress all the normal python
> alarms, found in
> http://svn.python.org/projects/python/trunk/Misc/valgrind-python.supp

Ok!

> If your MPI library also triggers errors, you might want to look for a
> suppressions file for that also. Alternatively, it doesn't take long to
> make your own. Run valgrind on a known-good test, with
> --gen-suppressions=all and put the generated suppressions into your .supp
> file. (You may want to copy just the first few lines of each, to make them
> more general, that way you will usually only have to copy a few.)

Nice! 

I suggest that we use the valgrind-python.supp together with a MPI suppression 
file and run the valgrind test on the python demos too.

Thanks for helping out with this. I am just a baby learing to crawl wrt 
valgrind.

> Another thing, the buildbots offer an enticing entry "memchecklog", but it
> seems to be always empty. What's up with that?
>
>
> And still happy, um, "romjul" to you all!

Thanks and to all of you other hacker too!

Johan